burnwatch 0.12.1 → 0.13.1

package/CHANGELOG.md

@@ -5,6 +5,32 @@ All notable changes to burnwatch will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.13.1] - 2026-03-25
+
+### Fixed
+
+- **OpenAI NaN spend / NaN% budget**: The OpenAI connector didn't guard against non-numeric `amount.value` responses. One NaN from OpenAI poisoned the entire brief total. Added type checking in the connector and a global NaN guard in `buildSnapshot()` — no single bad connector can ever corrupt the display again.
+- **Excluded services still showing in status**: Stripe, AWS, and other excluded services appeared in `burnwatch status` with ~$0.00 CALC. Now filtered out before polling — excluded services never appear in the brief.
+- **Anthropic shows BLIND despite having admin key stored**: The LIVE API call fails (endpoint/auth issue), and the CALC fallback requires `planCost` which was never set due to shell `$` variable expansion eating plan names like `"Max ($200/mo)"` → `"Max (/mo)"`. Three-part fix:
+  1. **Shell-safe plan matching**: Configure command now strips `$` amounts before fuzzy matching, so `"Max (/mo)"` correctly matches `"Max ($200/mo)"`.
+  2. **Registry plan fallback**: `pollService()` now resolves `planCost` from the registry plan's `monthlyBase` when `planCost` is missing but `planName` is set — so LIVE failures gracefully degrade to CALC instead of BLIND.
+  3. **Prorated CALC spend on failure**: When LIVE fails and falls back to CALC, spend is now prorated to day-of-month instead of showing $0.
+- **CALC services all showing $0 spend**: Flat-fee services (Vercel Pro, Browserbase Startup, etc.) had `planCost: 0` because the configure command's plan match failed silently. The shell-safe matching fix above prevents this for future configurations.
+
+## [0.13.0] - 2026-03-25
+
+### Added
+
+- **Utilization Engine** — Code-derived cost projection system that tracks SDK call sites across the project, persists them to `.burnwatch/data/utilization.json`, and projects monthly utilization and overage costs. This is the forward-looking early warning system: it tells you what your code *will* cost before you ship it.
+- **`burnwatch scan` command** — Full project scan for utilization patterns. Shows a table of services, call site counts, projected monthly units, plan inclusions, and overage costs. Use `--verbose` to see individual call sites per file.
+- **Incremental utilization tracking in hooks** — The PostToolUse hook now updates the utilization model after every file change. Overage warnings are injected into Claude's context when projected overage exceeds $5/mo.
+- **Utilization in session brief** — The SessionStart hook appends utilization data to the spend brief, showing projected usage alongside billing API data.
+- **LIVE vs utilization divergence alerting** — For services with billing API data (LIVE tier), the utilization model runs silently unless projected cost exceeds current spend by >50% AND >$5 (both thresholds required to avoid noise).
+- **Auto-scan on init** — `burnwatch init` now runs a full utilization scan after service detection. The utilization model is populated from day one.
+- **Utilization in interview JSON** — `burnwatch interview --json` now includes a `utilization` field with projected monthly units, plan inclusions, overage costs, and top call sites per service. Enables agents to say "your code projects 1,060 sessions/month, 960 over your plan."
+- **6 new `unitRate` values in registry** — Browserbase ($0.10/session), Resend ($0.001/email), Inngest ($0.01/fn run), SendGrid ($0.001/email), Replicate ($0.001/prediction). LLM services intentionally omitted (model variance too high for single rate).
+- **22 new tests** (70 → 92): Model operations, overage math, divergence alerting, file analysis, brief formatting, round-trip persistence.
+
 ## [0.12.1] - 2026-03-25
 
 ### Fixed